Algorithm 1: Velocity from Optical Flow.

Require: Video $V_1$, $V_2$, and $V_3$ from right, left, and back cameras with regions of interest $R_1$, $R_2$, and $R_3$. Ensure:$\dot{\theta }$ (angular velocity), $\dot{s}$ (linear velocity).  Procedure:    1. Let $t=1$ and acquire the first frame $I_1^i$ of $V_i$ in gray scale, $i=1,2,3$;    2. For each subsequent frame $I_{t+1}^i$ of input video, $i=1,2,3$:   (a) Acquire frame $I_{t+1}^i$ in gray scale.   (b) Detect sub-pixel accurate corners in $R_i$ for frame $I_t^i$ to obtain feature set $Q_I^i$.   (c) Calculate optical flow for $I_t^i$, $Q_I^i$, and $I_{t+1}^i$ to obtain corresponding sparse feature set $P_I^i$ using Lucas-Kanade.   (d) Remove keypoints without correspondences from $P_I^i$ and $Q_I^i$.   (e) Project $P_I^i$ and $Q_I^i$ from the image plane to the ground plane to obtain $P_G^i$ and $Q_G^i$ using the camera calibration information ($\mathtt{P}$ matrix) determined according to the method described in Section 3.1.   (f) Combine point set $P_G^i$ and $Q_G^i$, $i=1,2,3$, separately to obtain ground point sets $P_G$ and $Q_G$.   (g) Remove outliers from $P_G$ and $Q_G$ using RANSAC to obtain consistent ground plane point sets $P_G^{\prime }$ and $Q_G^{\prime }$.   (h) Compute rotation matrix $\mathtt{R}$ and translation vector $\mathbf{t}$ using $P_G^{\prime }$ and $Q_G^{\prime }$.   (i) Compute the rotational velocity $\dot{\theta }$ and linear velocity $\dot{s}$ using $\mathtt{R}$ and $\mathbf{t}$.  return $\dot{\theta }$ and $\dot{s}$.